Image Classification

Classification
Classification is the process where pixels are grouped according to the similarities in their spectral properties, that is, the range and distribution of reflected electromagnetic radiation captured by the Landsat TM sensor.  There are many different classification methodologies; the strategy that is adopted depends on the resources, type of ground truth and goals of the project.

In the case of Kentucky GAP, we have decided to use an iterated unsupervised classification strategy designed to provide a base layer for the vegetation layer modeling process. (Click on modeling for a description of this process.)

One of the limitations of unsupervised classification is that it is a spectral classification only.  That is, once the classification has been performed the classes produced do not necessarily correspond to a particular land cover type.  There is not a one to one correspondence between spectral classes and land cover.  However, the spectral signatures of certain cover types will tend to cluster around specific groupings of unsupervised classes. 

To reduce the spectral confusion between classes each general landcover type (wetlands, coniferous forest, deciduous forest, agriculture-urban-other) is separated from the others before a final set of spectral classes for that group is determined.  This is so that the spectral variability used by the classification algorithm is confined to that group of vegetation only. Classes that have similar spectral properties (e.g. forested wetland and coniferous) are maintained in separate data bases.

Wetlands
This landcover type is the most troublesome for classification purposes since the presence of water tends to confuse the spectral properties of the vegetation present there.  Although we are delaying in actually classifying the wetland areas until a suitable strategy can be devised, we do have a way of separating them from the rest of the vegetation types.  The National Wetlands Inventory (NWI) is a set of paper maps and corresponding digital files compiled by the U.S. Fish and Wildlife Service.  

NWI mapping for Kentucky was completed in 1988 and we are using those maps to delineate Kentucky's wetland areas.  With this information we can extract all wetlands from our images and proceed with classifying upland forested areas separately.  To do this we create a template layer based on the NWI polygons where all wetland areas have a value of 0 and all other areas a value of 1.  We then multiply the template image by the Landsat image.  Once multiplied, all wetland pixels have a value of 0 while all other pixels maintain their original values.  Separating wetlands is the first step in our classification procedure.

Coniferous Forest
The next step is to delineate, separate and classify all pixels containing coniferous vegetation.  Working with an image acquired during the winter (leaf-off) we first remove the wetland areas as described above.  A leaf-off image is used because coniferous vegetation is more easily identified visually and spectrally when the deciduous vegetation has no leaves.  The remaining pixels are subject to an unsupervised classification segmenting the pixels to 100 spectral classes. 

With   wetlands removed, the 100 classes still consist of coniferous, deciduous and all other pixels.   Once the image has been classified, National Aerial Photography Program (NAPP) 1:40,000 color infrared photos are used to help identify those unsupervised classes that contain coniferous vegetation, typically about 35 of the 100 classes.  This is the first iteration.

All pixels not contained within the 35 classes of interest are removed from further processing.  Then a second unsupervised classification is performed on the remaining pixels segmenting them to 100 new spectral classes.  This iteration is a way of refining the classification to include only those pixels that contain coniferous vegetation.  By reclassifying the original 35 classes to 100 classes the spectral variation present is segmented and we can refine the selection of pixels to those uniquely coniferous.  Pixels that were erroneously included in the first iteration are now removed.  This is the second iteration.

By the time the second iteration is finished, all pixels containing coniferous vegetation have successfully been separated from all other pixels.  This final group is then subjected to a third and final unsupervised classification of 100 spectral classes.  These classes contain only the spectral variation within the coniferous forest cover type and are used as a base layer in the modeling process.  A template of coniferous-only vegetation is prepared and used to remove these pixels from the Landsat images that will be used to classify the deciduous vegetation.

Deciduous Forest
This classification uses an image that has had both wetlands and coniferous vegetation removed before processing begins so that all that remains are deciduous cover types and everything else that is not coniferous or wetland.  The image is a composite of two leaf-on dates, spring and fall.  The theory behind using two dates is that phenological differences between different types of deciduous vegetation will be captured between the two dates.  This enhances the likelihood of successfully identifying different deciduous vegetation types.

An identical process of iterated unsupervised classification as described above is used to separate deciduous cover types from the remaining pixels in the image.  Once again, the final product is a base layer of 100 spectral classes that contain variation within the deciduous vegetation.  A template of deciduous forest pixels is created and used to remove these pixels from the image leaving only non-forested areas.

Agricultural - Urban - Other
What remains after wetlands, coniferous forest and deciduous forest have been identified and removed are agricultural lands, barren lands (strip mines, etc.) and urban areas.  Since these pixels have in effect, already been subject to two iterations of unsupervised classification, a third and final iteration is performed.  This group of classes contains many classes only of incidental interest to GAP and will be classified to Anderson Level I for urban classes, Level II for agriculture and barren lands including: Class 21 - cropland and pasture;   Class 24 - other agricultural land; Class 75 - strip mines; etc. (Anderson, J. et al, 1976, A Land Use and Land Cover Classification System for Use with Remote Sensor Data.   Geological Survey Paper 964).  These classes are also useful for defining ecotonal areas (boundary zones between ecotypes) and edges that are useful for wildlife mapping.